Hacker News Reader

I wonder what watermarks on an LLM output would look like. Say you have a draft for an email you have written and would like the LLM to correct typos and such. Would it be forced to change it substantially? If you tell it to reword it even so slightly, would the watermark still be valid?

View on HN

I think it’s a bunch of hidden UTF-8 character substitutions so the text looks the same but the underlying bytes are different

Thanks, that makes a lot of sense. I wonder however how can this be done without giving a lot of problems when copypasting or sending the input to other LLMs (not even thinking about code here).

I was wrong here, it’s more about the RNG they use for the word choices [1]. Although I wouldn’t be surprised to see UTF-8 substitutions also.

[1] https://www.anthropic.com/news/claude-text-watermark

Reply on news.ycombinator.com